In recent months, Amazon Web Services (AWS) has experienced at least two significant outages linked to the use of its AI coding tools. As a result, the company undertook a thorough review of how these assistants are integrated into their operations.
A report indicates that AWS, the largest cloud service provider, circulated an internal postmortem regarding these outages, particularly focusing on a system designed for customers to analyze and manage service costs. Following their investigation, Amazon determined that the involvement of AI tools was merely a “coincidence”, emphasizing that “such issues could arise from any developer tool or manual action.”
In a statement to the Financial Times (FT), Amazon classified both incidents as user-related errors, asserting that “the errors were due to user mistakes, not AI errors,” and stated their analysis found no indication that errors were more common when using AI versus traditional development methods.
The FT report describes a notable incident in December 2025, where a 13-hour disruption occurred after engineers allowed Kiro, Amazon’s AI coding assistant, to make changes autonomously. Following this, the tool concluded that the best solution was to “delete and recreate the environment.”
Amazon characterized this December incident as an “extremely limited event” that affected only one service in certain regions of mainland China. Additionally, the company clarified that the second outage did not disrupt any “customer-facing AWS services.”
Importantly, neither outage approached the severity of a separate 15-hour disruption in October 2025, which had taken down several customer applications and websites, including OpenAI’s ChatGPT.
Amazon noted continued customer adoption of Kiro, highlighting its commitment to enhancing efficiency for both users and employees. Following the December incident, “AWS implemented numerous safeguards,” the company stated, including mandatory peer review processes and enhanced staff training.
Feedback from Amazon Employees Regarding AWS Outages
Several Amazon employees shared insights with the Financial Times, noting that this was the second occasion within a few months where one of the company’s AI tools had been implicated in a service disruption. A senior AWS employee remarked, “We have already witnessed at least two production outages [in recent months]. The engineers allowed the AI [agent] to resolve an issue autonomously. While these outages were minor, they were entirely foreseeable.”
Employees indicated that the AI tools were treated similarly to human operators and granted equivalent permissions. In both incidents, the involved engineers did not seek second-party approval before making changes, a practice that is generally mandated.
Nonetheless, Amazon defended the Kiro tool, asserting it “requests authorization before taking any action” as a default measure. Furthermore, the company pointed out that the engineer involved in the December 2025 incident possessed “broader permissions than expected — a user access control issue, not an AI autonomy issue.”
Launched in July, AWS’s Kiro coding assistant is marketed as a significant advancement beyond “vibe coding,” enabling users to rapidly develop applications without writing extensive code by adhering to specific guidelines. Previously, the company relied on the Amazon Q Developer, an AI-enabled chatbot designed to assist engineers in coding. Witnesses reported that this earlier tool played a role in the prior outage.
Despite advancements, some Amazon employees expressed reservations about relying heavily on AI tools for critical tasks, citing potential risks of errors. They also mentioned that the company has set a goal for 80% of developers to utilize AI for coding tasks at least once weekly, actively monitoring adoption rates.
As AWS, responsible for 60% of Amazon’s operating profits, continues innovating in AI technology, including agents capable of executing instructions independently, it also aims to market this technology to external clients. The two recent incidents underscore the importance of careful implementation, as AI tools can sometimes produce unexpected results, leading to service outages.
